Graphics Reference
In-Depth Information
Figure
.
.
Estimated claim rates for selected values of Make and Zone
Conclusion
7.7
Wehave given four examples toillustrate the uses of GUIDE forbuilding visualizable
regressionmodels.Wecontendthat amodelisbestunderstoodifitcan bevisualized.
Butin ordertomake effective useofcurrent visualization techniques, namely, scatter
and contour plots, we will oten need to fit models to partitions of a dataset. Other-
wise, we simply cannot display a model involving more than two predictor variables
in a single
-D graph. he data partitions, of course, should be chosen to build as
parsimonious a model as possible. he GUIDE algorithm does this by finding par-
titions that break up curvature and interaction effects. As a result, it avoids splitting
a partition on a predictor variable whose effect is already linear. Model parsimony as
a whole is ensured by pruning, which prevents the number of partitions from being
unnecessarily large.
Ater pruning is finished, we can be quite confident that most of the important
effects of the predictor variables are confined within the one or two selected linear
predictors. hus it is safe to plot the data and fitted function in each partition and
to draw conclusions from them. As our examples showed, such plots usually can tell
us much more about the data than a collection of regression coe
cients. An obvious
advantage of
-D plots is that they require no special training for interpretation. In
particular, the goodness of fit of the model in each partition can be simply judged by
eye instead of through a numerical quantity such as AIC.
he GUIDE computer program is available for Linux, Macintosh, and Windows
computers from www.stat.wisc.edu/%
Eloh/.
